EP0273166A1 - Method and apparatus for preventing dead-locks in distributed data base systems - Google Patents

Abstract

Blocking events triggered in the individual nodes (K1 to K3) by subtransactions (STA...) ordered by a global transaction control are signalled to the global transaction control by blocking return message (SRM) which then informs all nodes in which subtransactions of the same global transaction (T...) are occurring, by means of blocking messages which leads to the associated subtransaction (for example STA 11, STA 21, STA31) being marked as "REMOTELY BLOCKED" in the respective local queuing table. Should a subtransaction of a global transaction run into a block due to a queuing graph (for example WG12) marked as "REMOTELY BLOCKED" or if a marking as "REMOTELY BLOCKED" is made for an existing queuing graph (WG22, WG32), a deadlock analysis (DLA) is triggered which is transferred step by step to the nodes concerned by means of the destination information associated with the blocking note until an open end of the global queuing graph formed by the local queuing graphs (for example WG12, WG22, WG32) via the respective row information is found or deadlock is determined. Extensive relief of the actual data processing by means of execution of these tasks by the control facilities for the distributed processing. <IMAGE>

Description

The invention relates to a method and an arrangement for avoiding global system deadlocks when processing global transactions in distributed database systems with several nodes according to the preamble of claim 1.

A number of technical problems arise in the implementation of distributed database systems. One of these problems is the synchronization of competing global transactions, in particular the avoidance of global deadlocks, because, for example, two or more transactions block each other waiting for the release of locked objects in order to be able to continue working - "Informatik-Spektrum", Springer Verlag, 1981, Issue 4, pages 175 to 184 and 261 to 270 or "Computing Serveys, Vol. 13 No. 2 June 1981, pages 149 to 183.

A whole series of processes have already been developed for this purpose. One of these methods consists of a kind of timeout monitoring in the respective origin node in that a transaction is simply canceled without a response after a predetermined waiting time has been exceeded, with each cancellation being associated with a reset of transactions and a repetition. With such a method, it cannot be ruled out that the time will only be exceeded because the current system load is too high without the risk of a system jam. The associated but unnecessary resetting of transactions and, in particular, the repetition caused by this increase the existing system load - see, for example, Informatik-Spektrum, page 182, section 3.3.1 or Computing Surveys, page 157, last paragraph of the left column .

The solution known from DE-OS 33 30 529.3 is based on the principle of the time lapse check, even if the monitoring is not carried out in the original node of a global transaction, but rather in the foreign node executing a partial transaction, provided that this partial transaction is there on a waiting partial transaction of another global transaction has accrued. This can reduce the likelihood that a transaction will be unnecessarily rolled back. However, an unnecessary reset cannot be ruled out either because the decision about it is also based on a presumption.

Other procedures are based on the allocation of time stamps. If a transaction encounters the lock of a younger transaction, the older transaction is rolled back. The constant admission of younger transactions means that the risk of a transaction being reset increases as the transaction ages. With the age and the number of conflicts, however, the size of a transaction increases, so that repetition becomes more and more complex and places an increasing burden on the system.

This contrasts with procedures in which a transaction is only rolled back if there is actually a deadlock. These procedures are based on checking the dependency given in the entire system using graphs, expediently by a central body. However, this requires an additional extensive and constant exchange of messages about the respective dependencies between the individual nodes and / or with the central monitoring point. The resulting load on the connecting line and the considerable effort for graph management are also disadvantageous.

The object of the invention is therefore to control competing global transactions so that global deadlocks can be clearly identified and transactions can only be reset without the monitoring by a central point and with very little, the connection lines between the nodes not noticeably burdensome message exchange there is actually a jam.

This object is achieved with regard to the method according to the invention by the features of claim 1.

Afterwards, a lock event when processing a global transaction means that all subtransactions not directly affected by the lock are treated as locked due to the lock messages and entered in the corresponding local waiting graph table if they have not yet been entered. Since these other subtransactions are inevitably in the dormant state, they always form the end link of the respective waiting graph, so that these waiting graphs are temporarily linked together to form a global waiting graph by the additional destination information about the node causing the block with the local waiting graph there that further messages about blocking relationships must be exchanged in the individual nodes. The performance of the responsible global tran This does not affect transaction control, since it is in the waiting state anyway due to the lock event.

Global deadlocks can only be expected if a local waiting graph is coupled together with at least one other to form a global waiting graph. This can be possible for two reasons, namely if the end link of an existing local waiting graph is to be marked as "REMOVED LOCKED", or if a subtransaction occurs on a waiting graph already marked as "REMOVED LOCKED". In both cases, therefore, a deadlock analysis is initiated, which is gradually expanded to the local waiting graphs of the nodes involved, which are linked in sequence by the destination, until an open end of the local waiting graph is reached or a deadlock is detected.

To carry out a deadlock analysis, a maximum of n messages must be sent to n affected nodes if the event triggering the deadlock analysis consists in the subsequent marking of an existing waiting graph as "EXTERNALLY BLOCKED" (closed deadlock analysis). In the event of a blocking event caused by a waiting graph already marked as "EXTERNALLY LOCKED", there are only n - 1 messages (linear jamming analysis), since the. Decision on the initiation of the deadlock analysis is based on the new entry element of the waiting graph.

In addition, a transmission of additional messages is only necessary if a deadlock is found to reset a transaction. However, this applies to all known processes.

The new solution principle is also based on a graph method with reliable detection of deadlocks. However, there is no need for a higher-level graph management of the system, and the number of messages loading the system is on average lower than in the known methods. The scope of the messages to be exchanged is also very small.

The latter also applies in particular to the information on the deadlock analysis orders, whereby according to the development according to claim 2, only two identity details are required for two global transactions, namely a constant as a check criterion and an identity that can be derived from the end link of the requesting node in order to find the associated waiting graph in the nodes to be controlled. The test is always limited to comparing the identity of the respective end link with the test criterion supplied. If there is a match, there is also a deadlock loop which must be broken in a known manner.

The triggering reasons for a deadlock analysis are only necessary conditions. They do not yet say that there is actually a deadlock. As a result, not every deadlock analysis that is initiated immediately will result in a deadlock being detected, especially not if the lockout event triggering the deadlock analysis is short-lived. According to the development of the invention according to claim 3, the lock feedbacks to the controlling global transaction control are therefore delayed by a predetermined period of time and a deadlock analysis to be initiated due to the same lock event is only released if the triggering lock condition still exists after the predetermined period of time. The system is also relieved of unnecessary messages in this way.

On the other hand, in the case of overlapping transaction orders of two global transactions that both run on a lock, it cannot be ruled out that two deadlock analyzes are started at the same time and an existing system deadlock is found simultaneously on different nodes, so that two transactions may be reset. This can be virtually excluded by the development of the invention according to claim 4.

The time stamps used here have a similarity to the time stamps in the known method described above. In contrast, the time stamps used here only serve as a comparison value for starting and stopping a deadlock analysis, but not for determining the sequence of the events occurring in the various nodes of the system. The clocks of the individual nodes therefore do not need to be synchronized with one another. Controlled by the time stamps, the triggering event with the larger time value in the loop direction always overruns the local time stamp values and overruns them in a closed waiting graph loop.

The latter method increases the scope of the messages by the additional time stamps and the memory requirement for storing the local time stamps per waiting graph. However, this additional effort goes hand in hand with avoiding the sending of unnecessary messages and resetting in fact only one transaction in the event of a jam.

According to claim 5, the cheapest locally blocked subtransaction is reset in the node in which the deadlock was detected, but not a subtransaction in another node, which would require additional effort for message transmission and a delay in resetting. Further advantages of the invention result from further developments in accordance with patent claims 6 to 9, which relate to the utilization of the control devices for the distributed processing for the various messages and messages, and for the monitoring of the occasions for initiating a deadlock analysis and for carrying it out, whereby the control devices work independently of the database control devices in that the necessary information is kept accessible by both in tables. In this way, the actual data processing within the distributed database system is only minimally impaired by the message exchange between the individual nodes required within the scope of the invention. Connection line failures also have no effect if, in the case of a non-releasable jamming analysis job, the marking as "REMOVED LOCKED" is canceled again during the subtransaction triggering the job and this is then placed back under normal local time monitoring to prevent permanent locks.

The arrangement according to the invention for carrying out the aforementioned method results from patent claim 10. It consists of a number of devices which are assigned to their tasks partly according to the respective global transaction control and partly to the respective node as a whole. These devices can be implemented in a known manner by means of memory-programmed sequences.

The actual data processing within the distributed database system remains the least impaired if, according to the development of claim 11, the devices that perform the individual control tasks and are assigned to the respective node as a whole are assigned to the control devices for the distributed processing as far as possible, thereby relieving the database control devices accordingly, separate order memories and the waiting graph table provided with access options both by the database control devices and by the control devices ensure largely independent work of the individual devices.

• FIG 1 ein Beispiel für einen sich über drei Knoten erstreckenden geschlossenen globalen Wartegraphen,
• FIG 2 ein Ablaufdiagramm zur Erläuterung des der Erfindung zugrundeliegenden Lösungsprinzips mit Bezug auf FIG 1,
• FiG 3 bis 5 Ablaufdiagramme in Anlehnung an das von FIG 2 beiVerwendung von zusätzlichen Zeitmarken für unterschiedliche gegenseitige Zeitverschiebungen der lokalen Uhren,
• FIG 6 eine Strukturübersicht für einen Knoten des Datenbanksystems,
• FIG 7 eine Übersicht der Verbindungsstrukturen von drei gemäß FIG 6 strukturierten Knoten mit einer globalen Transaktionssteuerung GTS in einem der Knoten,
• FIG 8A und 8B ein Ablaufdiagramm der globalen Transaktionssteuerung GTS von FIG 6 für die Bearbeitung von Transkationen,
• FIG 9 ein Ablaufdiagramm der Datenbanksteuereinrichtungen DBMS von FIG 6 für die Bearbeitung von Datenbankaufträgen der globalen Transaktionssteuerung,
• FIG 10 bis 14 Ablaufdiagramme der Kontrolleinrichtungen DCN von FIG 3 von FIG 3 für die Bearbeitung der verschiedenen Aufgaben gemäß dem der Erfindung zugrundeliegenden Verfahren,
• FIG 15 den schematischen Aufbau einer Wartegraphentabelle,
• FIG 16 den schematischen Aufbau einer Auftragstabelle für Sperr-Rückmeldungen und
• FIG 17 den schematischen Aufbau einer Auftragstabelle für Verklemmungsanalysen.

Details of the invention are explained below with reference to exemplary embodiments shown in the drawing. Show in detail

• 1 shows an example of a closed global waiting graph extending over three nodes,
• 2 shows a flow chart to explain the principle of the solution on which the invention is based with reference to FIG. 1,
• FiG 3 to 5 flow diagrams based on that of FIG 2 when using additional time stamps for different mutual time shifts of the local clocks,
• 6 shows a structure overview for a node of the database system,
• 7 shows an overview of the connection structures of three nodes structured according to FIG. 6 with a global transaction control GTS in one of the nodes,
• 8A and 8B show a flowchart of the global transaction control GTS from FIG. 6 for processing transactions,
• 9 shows a flowchart of the database control devices DBMS from FIG. 6 for processing database orders of the global transaction control,
• 10 to 14 flow diagrams of the control devices DCN from FIG. 3 from FIG. 3 for processing the various tasks according to the method on which the invention is based,
• 15 shows the schematic structure of a waiting graph table,
• 16 shows the schematic structure of an order table for lock feedback and
• 17 shows the schematic structure of an order table for deadlock analyzes.

1 shows the resulting waiting constellation in a distributed database system with three nodes K1, K2 and K3 if a global transaction GTA, which is designated T1 or T2 or T3, is carried out in each node under the supervision of the respective global transaction control. Each of these global transactions initially works with a sub-transaction STAU or STA21 or STA31 in its own node and blocks a data object OB there, namely objects A or B or C. Then the global transaction T1 wants to access the object with a second sub-transaction STA12 B in node K2, transaction T2 with a second subtransaction STA22 access object C in node K3 and transaction T3 with a second subtransaction STA32 access object A in node K1. However, since all three objects A, B and C are already blocked, the subtransaction STA12, STA22 and STA32 cannot be processed.

The three global transaction controls in the three nodes are thus waiting which can last forever because a system deadlock has occurred, which is externally recognizable by the concatenation of the arrows representing the individual subtransactions to form a closed sleep according to a closed global waiting graph. In order to be able to continue processing the data in the system, one of the subtransactions and thus the entire associated global transaction, for example the subtransaction STA31 and thus the global transaction T3, must be reset so that the system is operational again.

In order to keep distributed database systems operational, the risk of system jams must be recognized in good time. 2 shows a flow chart of the method ensuring this according to the invention when applied to the order constellation on which FIG. 1 is based.

The representation in turn captures the three nodes K1, K2 and K3 with the subtransactions STA11 or STA21 or STA31 which are currently being processed or are active. At any time, an order for the execution of the subtransaction STA12 from the node K1 to the node K2 is then issued, which requires the same object as the transaction STA21 there. The subtransaction STA12 therefore runs on a lock through the subtransaction STA21. Since the latter has not yet been entered in the local waiting graph table, the new waiting graph WG21 shown is entered in the waiting graph table. Furthermore, a lock confirmation SRM is sent back to the ordering node K1, which causes an entry of the new waiting graph WG11, consisting of the subtransaction STA11 with the marking as "REMOVED LOCKED" and with the destination K2 in the circle for the locking node in the waiting graph table there . The same processes take place in succession when the order for subtransaction STA22 is sent by node K2 to node K3 and when the order for subtransaction STA32 is sent by node K3 to node K1, so that ultimately the wait graphs WG12, WG22 and WG32 are created .

When transitioning from the waiting graph WG21 to the waiting graph WG22 in node K2 due to the lock feedback SRM from node K3, the one condition for initiating a deadlock analysis DLA is fulfilled for the first time since the end link of an already existing waiting graph WG21 is marked as "REMOVED LOCKED" and thus one global waiting relationship is established. This leads to an order to carry out a deadlock analysis at the node 3 designated by the destination with the identity of the global transaction T2 belonging to the end link STA21 as a search criterion for the associated waiting graph WG31 and the identity of the same Giobaitransaktio-T2 as a test criterion.

In node K3, the check of the waiting graph WG31 shows that the end link STA31 does not belong to the global transaction T2 and that the global waiting graph consisting of the local waiting graphs WG22 and WG31 has an open end. The deadlock analysis is therefore terminated without further effect.

Furthermore, the subtransaction STA32 in node K1 runs directly on the subtransaction STA11 marked as "REMOVED LOCKED" as a lock, so that the other trigger condition for a deadlock analysis DLA is given. In this case, the aim of the order for the deadlock analysis is node K2 with the identity of the Giobia transaction T1 derived from the subtransaction STAU for searching for the waiting graph WG22 in node 2 and the identity of the global transaction T3 derived from the subtransaction STA32 as triggering the event as a check criterion .

The check of the waiting graph WG22 in the node K2 triggered by this analysis job leads to a further deadlock analysis job at the node K3 with the identity of the global transaction T2 derived from the subtransaction STA21 as the end element as a new search criterion with unchanged check criteria. The check of the waiting graph WG32 triggered thereby in node K3 leads. finally to the end member of the global transaction T3 formed by the subtransaction STA31, whereby the proof for a closed loop in the checked global waiting graph is provided, so that a reset process which clears the deadlock is initiated.

Similar to the transition from the waiting graph WG21 to the waiting graph WG22 in node K2, the trigger condition for a deadlock analysis DLA is met in node K3 after the lock feedback SRM of node K1 arrives with the transition from the waiting graph WG31 to the waiting graph WG32 in node K3 the nodes K1, K2 and K3 are detected and the jam is detected in the node K3. In this case, instead of two, three orders have to be transferred for a deadlock analysis, since the starting point of the deadlock analysis is not the entry element of a waiting graph, such as the waiting graph WG12 in node 1, but the ending element of a waiting graph, e.g. in the case of the waiting graph WG32, in the node K3.

Irrespective of this, the specified order constellation leads to the same result, namely the determination of the system deadlock in node K3. Also, a transaction is only reset if there is actually a deadlock is present. The scope of the messages for the individual deadlock analysis orders is small and, in addition to the details of the type of order, only requires two identity details of transactions, one of which serves as a search criterion and the other as a check criterion.

By delaying the transmission of lock messages SRM by a predetermined period of time, provided that the lock condition is still present at the end of this time period, unnecessary lock feedback can be avoided if the triggering lock event is short-lived. The same applies to deadlock analysis jobs triggered by the same lock event if these are only sent to the target node after the lock replies SRM have been sent.

It cannot be easily avoided that messages for deadlock analysis jobs are unnecessarily transmitted, since analyzes that have been initiated can be terminated without result or triggered by a lock event in the system at two different nodes. In the present case, for example, at node K1 and subsequently due to the lock back SRM at node K3, although both analyzes lead to the same result.

However, order constellations are also possible in which two deadlock analyzes initiated simultaneously at different nodes due to mutually locked sub-transactions and intersecting lock feedbacks lead to the detection of a deadlock in two different nodes. In this case, two transactions are reset, but the reason is always an actual system deadlock.

Resetting only one transaction in the event of a deadlock and also reducing unnecessary messages for deadlock analysis jobs can be ensured by using additional time stamps as a time stamp. Each subtransaction entered in the local wait table is provided with a local time stamp, the value of which is derived from a local clock per node and indicates the time of the entry. In addition, each end link in the waiting graph receives a second time stamp for the entry of time values that are supplied as a time stamp by other nodes together with a lock message or an analysis job.

3 shows a flowchart corresponding to the process flow of FIG. 2 when using time stamps, assuming that all local clocks run synchronously, so that the deviation of the clocks at nodes K1 and K3 with respect to the clock at node K2 corresponds to At = 0 is zero.

If transaction T1 in node K2 is blocked by transaction T2, then the associated local time stamp receives the current value 1000 of the local clock as the blocking time t _sP . This value is also entered as the greatest value tLmax of the local time stamps of the waiting graph in the local time stamp of the blocking transaction T2, which is indicated by an arrow. However, the external timestamp of transaction T2 as the end link remains free for the time being.

The lock time specification t _s p ( _K2 ) is sent out with the lock feedback SRM and forwarded to the affected nodes with the lock message SM. In node K1, therefore, when transaction T1 is entered in the waiting graph table as "REMOVED LOCKED", the supplied locking time is entered in the external time stamp as the value t _Ex . In this case, the local time stamp remains free since there is no local blocking event for this waiting graph.

• 1. Der augenblickliche höchste Wert tLmax aller lokalen Zeit marken wird in die externe Zeitmarke übernommen und eingefroren, was durch einen Pfeil angedeutet ist.
• 2. Im Rahmen der Prüfung, ob eine der beiden Auslösebedingungen für die Verklemmungsanalyse erfüllt ist, wird gleichzeitig geprüft, ob die Sperrzeitangabe t_SP(K3) kleiner als der höchste Wert t_Lmax der lokalen Zeitmarken ist.

The same happens if the transaction T2 in node K3 at the time t _s p = 1025 is then blocked by transaction T3. In node K2, the associated lock feedback SRM does two things based on the lock time specification t _{SP (K3)} supplied:

• 1. The instantaneous highest value tLmax of all local time stamps is adopted in the external time stamp and frozen, which is indicated by an arrow.
• 2. In the course of checking whether one of the two triggering conditions for the _{deadlock analysis} is fulfilled, a check is carried out at the same time as to whether the blocking time _{specification} t _{SP (K3) is} less than the highest value t _{Lmax of} the local time stamps.

A deadlock analysis is only initiated if the last-mentioned condition is also fulfilled. In the present case, on the other hand, T _Lmax <t _SP ( _K3 ), so that, in contrast to the sequence in FIG. 2, no _{deadlock analysis} DLA is initiated.

As soon as transaction T3 from node K3 in node K1 directly arrives at the lock by transaction T1 marked as "REMOVED LOCKED" - lock time t _s p = 1500 - a local time stamp is created for the first time and therefore, as indicated by arrows, the blocking time t _s pin takes over the external time stamp. Further transactions that are _blocked in this graph change the local timestamp to the new value t _Lmax , but not the existing external timestamp. This will only be changed in the following if the marking as "REMOVED LOCKED" is removed again and subsequently a jamming analysis job or a new lock message is received etc.

In addition, as part of the check whether one of the two triggering conditions for the Ver Clamping analysis is fulfilled, now check whether the new blocking time t _{sP is} greater than the value of the external time stamp valid at the blocking time. In this case, this condition is met in accordance with _SP = 1050> t _Exalt = 1000, so that a deadlock analysis DLA is initiated, the ultimately valid value t _{EX of} the external time stamp, i.e. the value 1050, always being used for the test time t _DLA .

When the deadlock analysis is carried out in the controlled node, that is to say in node K2, an additional check is then carried out to determine whether the analysis test time t _{CLA is} greater than or at least equal to the value t _{Ex of} the external time stamp present. Only if this is the case is the analysis job forwarded to the next node with the original analysis test time t _DLA . In the present case, because of 1050 ≧ 1000, all conditions in node K2 are met, and the jamming analysis, which therefore extends to node K3, can determine the existing system jam there. In contrast, due to the lock feedback SRM at node K3, no deadlock analysis is initiated due to the unfulfilled time condition.

The flowchart of FIG. 4 differs from that of FIG. 3 in that, compared to the clock at node K2, the clock at node K1 runs according to At = + 100 and the clock at node K3 runs according to At = - 100. As a result of this, in contrast to FIG. 3, the lock-back avoidance SRM from node K3 to node K2 in node K2 leads to the initiation of a _{deadlock analysis} , since t _Lmax = 1000> t _s p ( _K3 ) = 925. The check then carried out in node K3, however, leads to the termination of the analysis since there is no end link associated with transaction T2 in the waiting graph. However, the supplied analysis test time t _DLA = 1000 is entered in the external time stamp there.

Otherwise, the time differences have no further effects on the rest of the process. When the lock feedback SRM arrives from node K1 in node K3, the existing value of the external time stamp is replaced by the highest value of the present local time stamp. In the flow chart of FIG. 5, the time shifts are reversed compared to that of FIG. The clock in node K1 runs ahead of the clock in node K2 according to At = - 100 and the clock in node K3 according to At = + 100.

In comparison to the procedure of FIG. 3, effects therefore only arise after the order for transaction 3 has arrived in node K1. Since the local blocking time t _s p = 950 is less than the value t _EXalt = 1000 of the existing external time stamp, no jamming analysis is initiated in node K1. Node K3 takes over this after arrival of the lock feedback SRM from node K1, since t _Lmax = 1125> t _{SP (K3)} ⁼ 950. The highest value t _Lmax = 1125 of the local time stamps is therefore frozen by transfer to the external time stamp and forms the analysis test time t _DLA . The search and test identity in this case both correspond to that of the end link which belongs to the global transaction T3. The deadlock analysis records all three nodes one after the other, because in node K1 and K2 the waiting graphs are transferred to the next node and the time condition t _DLA ≧ t _{EX is} fulfilled each time. The jam is then finally determined in node K3.

In comparison to the process sequence of FIG. 2, the use of time stamps completely suppresses the risk that two deadlock analyzes are triggered by a single blocking event or two transactions are added, except in the exceptional case that identical time stamps occur. As shown, the clocks at the individual nodes do not need to be synchronized, in contrast to the known time stamp methods.

6 shows the structural structure of a K ... node based on the SESAM-DCN database system from Siemens AG - publication U655-J1-Z55-2.

User programs ANW, which access local databases or databases in other computers as nodes, run under the supervision of a global transaction control GTS, which establishes the connection between the user programs and the database control devices DBMS of the desired database depending on the distribution rules defined in a database catalog. If the database is in its own node, access to the database control devices DBMS is direct. If the database is located in a foreign node, access is made in connection with the control devices for distributed processing DCN. The latter are responsible in particular for setting up the distribution rules, for accepting database accesses from other nodes and forwarding them to the corresponding database control devices DBMS, and for managing and monitoring the distributed database system. The connection of these devices of one node to the devices of another node is established via communication devices CAM with the connecting lines VL. Furthermore, the database control devices DBMS and the control devices DCN are coupled in the dedicated node K ... by buffer memory TAB, via which the orders are forwarded in one direction or the other can, so that database tasks and administration tasks can be processed largely independently of each other. Above all, these buffer memories also include the local waiting graph table, which can be controlled by both sides and changed if necessary.

The resulting traffic relationships in a distributed database system with three nodes between a global transaction control GTS at node K1 and the databases in the three nodes K1 to K3 can be seen from the block diagram of FIG. The global transaction control GTS in node K1 is directly connected to the database control devices DBMS at its own node in order to carry out data access and to receive results of the data processing, with direct control and modification of the local waiting graph table being possible. Data accesses TA-AUF to remote database control devices DBMS2 and DBMS3 at the other nodes K2 and K3, on the other hand, run via the local control devices DCN2 and DCN3, while results TA-ERG are each sent back directly from the database control devices DBMS2 and DBMS3 to the ordering global transaction control GTS. Monitoring messages M that are also required are exchanged directly between the respective global transaction control GTS and the control devices DCN2 or DCN3 mentioned. In addition, all control devices DCN1 to DCN3 are connected to one another for administration purposes.

In the case of distributed database systems with global transaction controls GTS at several nodes, the corresponding nodes would have to be supplemented with a global transaction control with their traffic relationships in the same way as the node K1 in FIG.

As is known, the transactions to be carried out by a global transaction control GTS consist of a sequence of instructions which are taken from an execution plan. Each transaction is opened with a start instruction BTA, whereby, in accordance with the underlying two-phase lock protocol, a required data object is locked and the lock is confirmed before a transaction can access the object, and all objects required by a transaction are only released again when the transaction can be ended or needs to be reset. A two-phase end protocol is used to end a transaction with the end statement ETA, after which only after a phase 1, which indicates the provisional end of the transaction, with a phase 2 for the final end, the database changes are committed and that System is brought into the next consistent state. If, on the other hand, a transaction cannot be completed, it is reset with a reset instruction RTA.

This assuming that the operation of the global transaction control GTS is roughly described with reference to the flowcharts of FIG. 8A and FIG. 8B, FIG. 8A referring to the execution of the instructions in the present execution plan when executing transactions and FIG. 8B to the reaction in accordance with incoming feedback relates.

With the start of the global transaction control GTS, a first order is taken from the execution plan. If the end of the orders for a transaction has not yet been reached, the database catalog is used to determine the target node in which the required database is located and a subtransaction list is used to check whether a subtransaction already exists for the target system. If there is no subtransaction yet, this is entered in the subtransaction list and a BTA instruction with an order is transferred to the target node. Otherwise, only the new order for the existing subtransaction will be sent.

When the end of a transaction is finally reached, ETA instructions are sent to all target nodes affected by a subtransaction and the current transaction is ended. After placing one of the addressed orders, the work of the global transaction control GTS is stopped for the current transaction. The transaction control is in the waiting state for feedback that determines the progress of the work.

The recipients of all the orders are the respective database control devices DBMS in the respective target node, the control of the database control device DBMS in foreign nodes taking place in each case via the control devices DCN there, while the database control devices DBMS are controlled directly in the own node and by the control devices DCN in the foreign nodes some of the tasks to be carried out here are carried out and carried out by the local global transaction control GTS.

FIG. 8B shows the progress of the workflow of the global transaction control GTS after the arrival of two responses which have an effect within the scope of the present invention. If the control devices DCN of a foreign node or the local database control devices DBMS report that a transaction order has been blocked, then the associated subtransaction is shown in the existing subtransaction list as marked locked. It is also checked whether further subtransactions of the same global transaction are listed in the subtransaction list. If the latter is not the case, then the global transaction control returns to the waiting state. If, on the other hand, further subtransactions are specified in the subtransaction list, blocking messages SM are sent to the control devices DCN of all associated foreign target nodes with the order to mark the corresponding subtransaction in the waiting graph table WTB there as "REMOVED LOCKED", while for the corresponding identification one on one's own The local waiting graph table WTB is addressed directly by the global transaction control GTS and the marking is made.

Due to a lock-back confirmation SRM according to FIG. 2 to the global transaction control GTS, all affected target nodes are informed of this lock event under the supervision of the global transaction control, directly and / or with separate lock messages SM, which are evaluated by the responsible control devices DCN. If a confirmation of execution arrives in the further course of the transaction order which has temporarily been blocked, then in the case of a blocked note in the subtransaction list, this is deleted again, a new transaction order is started and thus transferred to the process of FIG. 8A.

Database orders coming from the respectively controlling global transaction control GTS to the database control devices DBMS as a controlled database system result in the workflow shown in the flow chart of FIG. 9. First, it is checked whether there is an end or reset order ETA or RTA. In this case, the pending subtransaction must be completed by committing the changed or unchanged status and then releasing the locked data objects. If there are entries in the waiting table WTB for the transaction to be completed, these are deleted and the associated waiting graph updated. The execution of the order is then acknowledged in a known manner to the ordering global transaction control GTS.

If a pending subtransaction could not be ended, but a new order has to be executed, it is first checked whether the required data object is free at all. If this is free, it will be blocked. If necessary, the waiting graph table has to be adjusted and an existing entry in the order table LOCK-TAB has to be deleted if the order was previously locked and the object was only subsequently released again. The order is then carried out in preparation and the result is communicated to the ordering global transaction control GTS with confirmation of execution.

If, on the other hand, the desired object is not free but blocked by another transaction, the new order is blocked by this other transaction. This locking event therefore leads to an entry in the waiting graph table WTB and thus to a classification in a local waiting graph of the waiting graph table. At the same time, an order for a global transaction control GTS is entered into an order table LOCK-TAB on the basis of the lock event and thus in the effect of the invention. Subsequently, in a further effect of the invention, it is to be checked whether a trigger condition for a deadlock analysis DLA is fulfilled. For this purpose, as already explained, the end link of the associated waiting graph of the waiting graph table is checked. If this end link is not marked as "REMOVED LOCKED", there is no global waiting relationship in the system and it can be switched to a waiting state until the desired data object becomes free.

If, on the other hand, the end link is marked as "REMOVED LOCKED", then when using time stamps according to the one method variant according to the invention, it must be checked whether the time condition for triggering a jamming analysis DLA is also fulfilled. In the event that the conditions to be met are met, a corresponding analysis order is entered in an order table DLA-TAB.

Regardless of the outcome of these additional checks, the aforementioned waiting state is also assumed at the end and the release of the blocked object is monitored.

In the present case, it is assumed that the check for the existence of a trigger condition for a deadlock analysis DLA is carried out by the database control devices DBMS itself. This is obvious since the waiting graph to be checked is to be found in the waiting graph table WTB anyway because of the lock entry to be made. This task can, however, be transferred to the local control devices DCN in the same way as the sending of the lock-back report SRM using the order table LOCK-TAB and the jamming analysis orders using the order table DLA-TAB, by simultaneously with the entry in the order table GTS-TAB in one a corresponding order is stored in another order table or by triggering the check as a result of a direct call by the database control devices, so that the latter are exonerated from this test.

The basic structure of a waiting graph table WTB is shown in FIG. 15. For each transaction to be entered in a waiting graph, an entry line with the associated identity TA-ID is provided. Several lines can be chained to follow-up addresses F-AD in a manner known per se to form a waiting graph. Each line also contains display fields, namely STATE to identify the current status of the transaction as ACTIVE, LOCKED or REMOVED LOCK, TARGET to identify the node of the removed lock, TIME _L as a local time stamp to identify the local lock time and TIME EX as an external time stamp.

In the present example, the entries relate to a waiting graph with the transactions T1, T3, and Tx, the transaction T1 representing the end link and being marked as "REMOVED LOCKED" in node K2. Transactions T3 and Tx subsequently accrued on this transaction T1 and are therefore marked as "LOCKED". The local timestamps contain the respective time of the entry, the timestamp of the end link in each case containing the highest value of all local timestamps of the waiting graph. In contrast, the external time stamp takes on a time value triggered by external events, which has already been described with reference to FIGS. 3 to 5.

The essential tasks within the scope of the method on which the invention is based are carried out by the individual control devices DCN with a few exceptions at the node of the global transaction control GT which is the ordering party. These are the sending of lock feedback SRM from foreign nodes to the respective controlling global transaction control GTS, the marking of a transaction as "REMOVED LOCKED" based on a lock message SM of the controlling global transaction control GTS in foreign nodes with a check for a trigger condition for the deadlock analysis DLA, the allocation of analysis orders based on the order table DLA-TAB, the implementation of a deadlock analysis DLA based on a received analysis order, possibly with the forwarding of the order to the continuing global waiting graph or with determination of the transaction to be reset in the event of a detected deadlock and the transfer of a corresponding reset order to respectively associated local database control devices DBMS.

The corresponding work processes are explained in more detail below with reference to the flow diagrams from FIG. 10 to FIG. 14, it being assumed that these tasks in connection with the other tasks of the control devices DCN are addressed repeatedly as part of a cyclically working superordinate task management and that the respective orders are executed.

Regarding the sending of the lock feedback SRM from foreign nodes to the respective ordering global transaction control GTS according to FIG. 10, it is assumed that during a connection between the global transaction control GTS and the respective database control devices DBMS in a foreign node, the associated control devices DCN are periodically Intervals send a control message TACK back to the global transaction control GTS to indicate that the connection exists. In this case, the control messages TACK can be supplemented according to a development of the invention by additional information on the lock feedback SRM.

After calling the task "SEND CONTROL TACK TACK", the order table LOCK-TAB is therefore first checked for the existence of an order for a lock response to the global transaction control GTS to be notified. If there is no order, the control message TACK is sent in the usual form. If, on the other hand, a lock confirmation is to be sent to SRM, the associated information for the lock feedback is taken from the order table LOCK-TAB and a control message with lock information is set up and sent.

On the other hand, in systems in which connection failures are monitored without the repeated sending of TACK control messages, the lock-back messages are transmitted as independent messages.

The removal of the order from the order table LOCK-TAB. what is associated with a deletion of the entry takes place in the present example with a delay by the time period AT, that is only if, at the time of access, the value of the local clock is at least equal to the lock time indication t _AP for the lock feedback, which is increased by the value AT SRM is. Therefore, if entries in the order table LOCK-TAB are deleted by the database control devices DBMS according to FIG. 9 before the waiting time AT expires because the lock has now been dispensed with, unnecessary lock-back messages are omitted, which could possibly trigger the transmission of unnecessary messages for carrying out jamming analyzes .

The basic structure of the order table LOCK-TAB is shown in FIG. 16. An entry includes the identity GTS-ID of the global transaction control GTS to be controlled, the identity TA-ID of the transaction concerned, the destination specification ORT sp for the blocking notice to be undertaken and the blocking time stamp ZEIT _SP with the blocking time specification tsP.

The workflow of FiG 11 shows the reaction of the control device DCN in a foreign node after receipt of the lock message SM, which results in an order to mark a transaction as "REMOVED LOCKED". Based on this order, the waiting graph table WTB is first checked and it is determined whether the transaction concerned has already been entered. If this is not the case, a corresponding new waiting graph is created, consisting of the defident TA-ID of the transaction concerned, the status information "REMOVED LOCKED" and the destination of the remote lock. If additional time stamps are used, the specified blocking time specification t _{S (K ...)} is also entered in the associated external time stamp ZEIT _EX . - See Ze4e 1 of the waiting graph table WTB of FIG. 15.

If, on the other hand, the transaction in question is already contained in the waiting graph table WTB, the associated waiting graph is supplemented in the same way. On the other hand, one of the triggering conditions for checking whether a DLA deadlock analysis should be initiated is met. When using time stamps, it is also checked whether the specified time conditions are met. If the latter is the case, an entry is made in the order table DLA-TAB for carrying out a jam analysis.

The basic structure of this order table DLA-TAB is shown in FIG 17. In this case, an entry includes the identity DCN-ID of the control devices to be controlled, the identity TA-ID _{ON of} a transaction as a search criterion for the waiting graph concerned, the identity TA-ID _{DLA of} a transaction as a check criterion for checking for the presence of a jam, and when using time marks ZEIT _DLA the value of the analysis test time t _DLA .

As already mentioned, when a local transaction is marked at the location of the global transaction control GTS, the task described with reference to FIG. 11 is carried out by the global transaction control itself as REMOVED LOCKED.

The deadlock analysis DLA to be initiated on the basis of an entry in the order table DLA-TAB, on the other hand, is carried out solely by the control devices DCN of the nodes concerned, without involving the controlling global transaction control GTS. FIG. 12 shows the workflow with regard to the issue of a jam analysis order and FIG. 13 shows the workflow with regard to the execution of the job in the received node.

As soon as the task "CHECK DLA TABLE" is called up by the task management in one of the control devices DCN, it is checked whether there is an entry in this. If there is an order, the information for this is taken and the message for a jam analysis job is created. Then, for example, the connection to the desired destination node is established and, as soon as it is available, the job is sent. If, on the other hand, a connection cannot be established and the job cannot be forwarded, a connection failure must be assumed. In this case, the lock note of the associated transaction is deleted in the waiting graph table and marked as "ACTIVE". The transaction is again subject to the usual local time monitoring to prevent permanent locks and can be reset after the time has expired. The connection failure thus has no effect. The same applies to systems in which messages are sent immediately without prior checking of the actual establishment of the connection and the receipt of each message is acknowledged as soon as there is no acknowledgment.

The "ORDER TO CARRY OUT A CLAMPING ANALYSIS" arriving at the target hotspot is then processed in the manner represented by FIG. 13. It will first be examined on the basis of the _S uchkriteriums the associated wait graph in the Wartregraphentabelle WTB and checks the end of this waiting graphene using the test criterion. If the end link belongs to another global transaction, it is checked whether there is a mark as "REMOVED LOCKED". If this is not the case, the analysis can be terminated. Otherwise, the job must be forwarded to the new target node by an entry in the job table DLA -TAB is made.

If additional time stamps are used, the forwarding of the analysis job is also dependent on compliance with the necessary time conditions.

If, on the other hand, the end link of the checked waiting graph belongs to the global transaction sought, there is a system deadlock. A suitable subtransaction for resetting is therefore determined on the basis of the waiting graph and a corresponding order is entered in the order table RES-TAB.

As part of a new task assignment "CHECK RES-TAB" it is then checked whether there is a reset request. If this is the case, the associated data are extracted and a reset message for the responsible local Database control devices DBMS set up and handed over to them. According to FIG. 9, these then cause the subtransaction named in the order to be reset and report this directly to the responsible global transaction control, which then uses the subtransaction list to send the further reset requests to the target nodes concerned. If a reset request cannot be accepted by the DBMS responsible for any reason, the waiting graph table WTB with the identity of the transaction to be reset is looked up in the waiting graph table WTB and the blocking note of the associated end link is deleted and marked as "ACTIVE" so that also in this case it is transferred back to the local time monitoring.

The method explained with reference to FIG. 2 to FIG. 5 is thus fully integrated into the control of a distributed database system, the actual data processing being loaded only insignificantly by the individual database control devices DBMS of the distributed system, since all essential procedural tasks are performed by the control devices DCN of the system . The relevant global transaction control GTS has the primary task of accepting lock-back reports and, derived therefrom, of forwarding lock-ups to the nodes affected by subtransactions of the same global transaction. The global transaction control then carries out additional tasks resulting from a lock event only in connection with the local database control devices DBMS in its own node. In addition, the control control devices DCN are responsible at the individual nodes, in particular for the forwarding of analysis orders and their execution.

Claims (12)

1.Procedure to avoid global deadlocks when processing global transactions (e.g. T1 to T3) in distributed database systems with several nodes (e.g. K1 to K3), the user processes running in each of the nodes being handled by a global transaction control (GTS) and where each node with local databases database control devices (DBMS) for the assignment and lifting of locks on local data objects, for checking and managing local waiting relationships of current transactions or independent parts thereof as partial or sub-transactions (STA) and for resetting transactions and each node is equipped with control devices (DCN) for the distributed processing and monitoring of existing connections and with devices (CAM) for exchanging messages between the individual nodes, characterized in that - that each sub-transaction order from a global transaction control (GTS) and a lock triggers a lock response (SRM) to the associated global transaction control (GTS) and this for all other sub-transactions of the same global transaction (e.g. T1) lock messages (SM) together sent with a destination information about the node triggering the block (eg K2) to all affected nodes to identify the subtransaction pending there with a block note as "REMOVED BLOCKED", - that each sub-transaction that has been blocked in one of the nodes (e.g. K1) and that has not yet been entered due to a lock message (SM) and has not yet been entered with a lock note is entered in the local waiting graph table (WTB) of the respective node (K1) in in the case of a blocking message (SM), the blocking note and the specified destination information are also entered, -that a deadlock analysis (DLA) is initiated when a blocking note is entered for the end link of a waiting graph (WG) already contained in the waiting graph table (WT8) or when a subtransaction is entered in a waiting graph whose end link is already marked as "REMOVED LOCKED" by a locking note which is gradually extended to the nodes (e.g. K2, K3) whose local waiting graphs (WG) are temporarily connected to a global waiting graph by the destination information associated with the blocking information, until an open end of the global waiting graph is found or a deadlock is detected which is then canceled by resetting an appropriate transaction. 2. The method according to claim 1, characterized in that for carrying out a deadlock analysis (DLA) two identity details (for example T1, T3) are provided, one of which (for example T3) as a uniform test criterion for the detection of a deadlock in all of the Depending on the event triggering the deadlock analysis (DLA), the nodes involved (e.g. K2 and K3) are equal to the identity (T3) of the global transaction, the subtransaction of which has accumulated on a waiting graph (WG) marked as "REMOVED LOCKED", or the subtransaction afterwards in a waiting graph (WG) is marked as "REMOVED LOCKED", while the other (T1) is a search criterion for finding the the corresponding waiting graph in the respective node (e.g. K2 or K3) is equal to the identity (T1) of the global transaction, the subtransaction of which is marked as "REMOVED LOCKED" in the respective node (e.g. K1), and that a deadlock is found to exist if in any of the nodes (eg K3) detected by the deadlock analysis (eg K3), a subtransaction marked as "REMOVED LOCKED" belongs to the global transaction (T3) specified by the check criterion. 3. The method according to claim 1 or 2, characterized in that the transmission of the lock response (SRM) is delayed for a predetermined period of time in a subtransaction order that has been blocked, that the lock feedback (SRM) is suppressed if during this period the lock condition is no longer present, and that a deadlock analysis (DLA) to be initiated on the basis of the lock condition is only initiated when the lock feedback has been sent. 4. The method according to any one of claims 1 to 3, characterized in that all transactions or subtransactions entered in a waiting graph table (WTB) are provided with a local time stamp (TIME _L ) derived from a local clock, -that this timestamp is added as a blocking time specification (T _{SP (K ...)} ) to the blocking feedback (SRM) and the blocking messages (SM) triggered by it, -that when a blocking message (SM) arrives in the case of a waiting graph that does not yet exist in the waiting graph table (WTB), a new waiting graph consisting of the locked blocked subtransaction is created and given an external time stamp (ZEIT _EX ) into which the supplied blocking time specification ( t _{SP (K ...)} ) is entered, while the largest value (t _Lmax ) of the existing local time stamps (TIME _L ) of the waiting graph is transferred to the external time stamp (ZEIT _EX ) if a waiting graph has already been created, - that, on the other hand, in the event of a block triggered for the first time by a subtransaction already marked as "REMOVED LOCKED", the value of the associated local time stamp is adopted as a new value in the external (ZEIT _EX ), - That a deadlock analysis (DLA) is only triggered if the value (t _s p) of the new local time stamp (TIME _L ) is greater than the value (t _EXalt ) of the in the event of a blocking event triggered by a subtransaction already marked as "REMOVED LOCKED" external time stamp to be renewed (TIME _Ex ) or if the waiting graph to be marked as "REMOVED LOCKED" later, the greatest value (t _Lmax ) of the existing local time stamps (TIME _L ) of the waiting graph is greater than the reported locking time specification (t _{SP (K. ..)} ) is - that when a deadlock analysis (DLA) is initiated, the last valid value (t _{EX new} ) of the external time stamp (ZEIT _EX ) of the waiting graph triggering the analysis job is added as a time stamp (T _DLA ) and - that the initiated deadlock analysis (DLA) in the affected nodes (e.g. K2 or K3) is only continued if the value of the time stamp (T _DLA ) that is also included is greater than or equal to the value (t _ex ) of the external time mask existing in the respective node ( TIME _EX ) is. 5. The method according to any one of claims 1 to 4, characterized in that in order to cancel a detected deadlock, the cheapest locally blocked subtransaction in the node (e.g. K3) in which the deadlock has been detected is reset. 6. The method according to any one of claims 1 to 5 for distributed database systems with monitoring of the existing connections between the individual nodes by repeated control feedback (TACK) or acknowledgments, characterized in that orders for lock feedback (SRM) from the local database control devices (DBMS ) are entered in a table (LOCK-TAB) which is checked from time to time by the control devices for distributed processing (DCN) when a control message (TACK) is to be sent to a global transaction control (GTS), and that at If there is a valid order for a block confirmation (SRM) to this global transaction control (GTS) of the control message (TACK), the block information corresponding to the order is added. 7. The method according to any one of claims 1 to 6, characterized in that from a global transaction control (GTS) lock messages (SM) evaluated by the control devices for distributed processing (DCN) in the receiving destination node and the necessary entries in the local waiting graph table ( WTB) can be done directly by them. 8. The method according to any one of claims 1 to 7, characterized in that orders for deadlock analyzes (DLA) are each entered in a table (DLA-TAB) which is checked from time to time by the control devices for distributed processing (DCN), that when an order is available for another node, the forwarding of a corresponding message is initiated, that the deadlock analysis is terminated in the case of a non-transferable message and the subtra that triggered the order and is marked as "REMOVED LOCKED" in the local waiting graph Action is marked directly as "ACTIVE", and that orders received by the control devices for detailed processing (DCN) in the controlled node (for example K2) for carrying out a deadlock analysis (DLA) are carried out by the latter directly by checking the local waiting graph table (WTB) In the event of a jam, a suitable transaction is determined and a reset request derived therefrom is entered in a table (RES-TAB), from where it is forwarded to the responsible database control devices (DBMS). 9. The method according to claim 8, characterized in that in the event that an existing reset job is not accepted by the associated database controls (DBMS), the waiting graph containing the transaction to be reset is looked up in the waiting graph table (WTB) and that is identified as "REMOVED LOCKED". marked associated end link is marked as "ACTIVE". 10. Arrangement for performing the method according to one of claims 1 to 9 with waiting tables (WTB) for displaying the local waiting relationships of locally pending transactions or sub-transactions, characterized in that each global transaction control (GTS) - facilities for receiving blocking feedback (SRM) from external nodes (eg K2 and K3), -Facilities for monitoring blocking events at your own node (e.g. K1), Equipment for the determination of the remaining subtransactions that of _'a barrier-feedback (SRM) affected global transaction, Devices for generating and issuing corresponding blocking messages (SM) to external nodes (for example K2 and K3), -Devices for checking and changing the entries in the waiting graph table at your own node (eg K1) depending on received lock feedback messages (SRM) or depending on lock events and occurred on your own node (eg K1) -Devices for checking the existence of a related trigger criterion for a deadlock analysis using the local waiting graph table (WTB)
and that each node in question (eg K1 to K3) is independent of any global transaction control (GTS) that may exist on the node - Devices for generating and sending block feedback (SRM) to the relevant global transaction control (GTS) depending on sub-transactions that have been blocked, - facilities for receiving blocking reports (SM), -Devices for checking and changing the entries in the waiting graph table (WTB) depending on received locking messages (SM) or locking events that have occurred and for checking for the existence of an associated trigger criterion for a deadlock analysis using the waiting graph table (WTB), Devices for generating and sending deadlock analysis jobs and for receiving deadlock analysis jobs, -Devices for checking for a system deadlock depending on received deadlock analysis orders using the waiting graph table (WTB) and -Devices for generating reset jobs when a system jam is detected. 11. The arrangement according to claim 10, characterized in that the devices provided in the nodes (for example K1 to K3) independently of the possibly existing global transaction control (GTS) are assigned to the control devices (DCN) for the distributed processing as sub-devices, that both the Database control devices (DBMS) as well as the corresponding sub-devices of the control devices (DCN) for each node (eg K1) are coupled to the local waiting graph table (WTB) and can access them depending on each other and that for the storage of orders for lock feedback (SRM), Buffer memories (LOCK-TAB, DLA-TAB, RES-TAB) are provided for carrying out deadlock analyzes (DLA) and for resetting transactions (RESET). 12. The arrangement according to claim 11, characterized in that the buffer memory (DLA-TAB, RES-TAB) for jam analysis and reset jobs are designed as a FIFO memory.

Images